正在确认阅读权限…
How Speech Recognition Turns Sound into Text
语音识别如何把声音变文字
精读先不看文字，完整听一遍。
-
Microphones convert sound waves into electrical signals, which analog-to-digital converters sample thousands of times per second.麦克风将声波转换为电信号，模数转换器每秒采样数千次。
-
Acoustic models break audio into tiny frames — typically 10-millisecond windows — and classify each frame’s phonetic likelihood using neural networks.声学模型将音频切分为微小帧（通常为10毫秒窗口），并用神经网络对每帧的音素概率进行分类。
-
Language models predict probable word sequences based on grammar, context, and vast text corpora, helping resolve ambiguities like 'there' vs. 'their'.语言模型基于语法规则、上下文及海量文本语料，预测可能的词序列，从而解决‘there’与‘their’等歧义问题。
-
Speaker adaptation techniques adjust recognition for individual voices, accents, or background noise using recent utterances.说话人自适应技术利用近期语音片段，针对个体嗓音、口音或背景噪音调整识别效果。
-
End-to-end systems now skip intermediate steps, mapping raw audio directly to text with attention-based transformers trained on millions of hours of speech.端到端系统跳过中间步骤，直接将原始音频映射为文本，依靠基于注意力机制的Transformer模型，训练数据达数百万小时语音。
-
Real-time transcription requires buffering short segments, predicting words before full sentences finish, and correcting errors on-the-fly.实时转录需缓存短音频片段，在完整句子结束前预测词语，并即时修正错误。
-
Privacy-conscious devices process voice locally whenever possible, sending only anonymized snippets to servers for improvement.注重隐私的设备尽可能在本地处理语音，仅向服务器发送匿名化片段以优化系统。
-
Accuracy exceeds 95% in quiet settings but drops significantly with overlapping speakers, heavy accents, or technical jargon without custom training.安静环境下准确率超95%，但多人重叠说话、浓重口音或未经定制训练的专业术语会显著降低准确率。
精读 · 大约 10 分钟
听 → 读 → 学表达 → 懂 → 今天会说这一句 → 跟读录音 → 自己说一句。
③ 学 · 重点表达
- Speaker adaptation techniques adjust recognition for individual voices, accents, or background noise using recent utterances. 说话人自适应技术利用近期语音片段，针对个体嗓音、口音或背景噪音调整识别效果。
- Privacy-conscious devices process voice locally whenever possible, sending only anonymized snippets to servers for improvement. 注重隐私的设备尽可能在本地处理语音，仅向服务器发送匿名化片段以优化系统。
- Microphones convert sound waves into electrical signals, which analog-to-digital converters sample thousands of times per second. 麦克风将声波转换为电信号，模数转换器每秒采样数千次。
④ 懂 · 选一选
⑤ 今天会说这一句
Speaker adaptation techniques adjust recognition for individual voices, accents, or background noise using recent utterances.
⑥ 跟读并录音
Speaker adaptation techniques adjust recognition for individual voices, accents, or background noise using recent utterances.
第一版：录下来、回听、再听标准音。系统打分以后再加。
⑦ 说 · 自己来一句
用今天的表达，自己说一句。
例子：Speaker adaptation techniques adjust recognition for individual voices, accents, or background noise using recent utterances.
已记下。连续天数以后会显示在课表里。
陈老师指导
课程开发中，敬请期待。成人和孩子都可以先自学听读、精读。
这边的讲解音轨还在做。想先跟着老师走完固定 30 课，可以去陈老师带读社（和自学通行不是同一份会员）。
课程开发中，敬请期待。